NOT F O R QUOTATION WITHOUT P E R M I S S I O N O F T H E AUTHOR
D E S I G N I N G APPROXIMATION SCHEMES FOR S T O C H A S T I C O P T I M I Z A T I O N PROBLEMS,
I N P A R T I C U L A R F O R
S T O C H A S T I C PROGRAMS WITH RECOURSE
J o h n B i r g e R o g e r J - B W e t s N o v e m b e r ' 1 9 8 3 W P - 8 3 - 1 1 1
W o r k i n g
P a p e r s a r e i n t e r i m r e p o r t s o n w o r k of t h e I n t e r n a t i o n a l I n s t i t u t e f o r A p p l i e d S y s t e m s A n a l y s i s and have received o n l y l i m i t e d r e v i e w . V i e w s o r o p i n i o n s e x p r e s s e d h e r e i n do n o t n e c e s s a r i l y repre- s e n t t h o s e of t h e I n s t i t u t e o r of i t s N a t i o n a l M e m b e r O r g a n i z a t i o n s .I N T E R N A T I O N A L I N S T I T U T E F O R A P P L I E D S Y S T E M S A N A L Y S I S A - 2 3 6 1 L a x e n b u r g , A u s t r i a
PREFACE
The System and Decision Sciences Area has been involved in procedures for approximation as part of a variety of projects involving uncertainties. In this paper, the authors discuss approximation methods for stochastic programming problems. This is especially relevant to the Adaptation and Optimization project since it directly applies to the solution of optimization problems under uncertainty.
Andrzej P. Wierzbicki Chairman
System and Decision Sciences Area
DESIGNING APPROXIMATION SCHEMES FOR STOCHASTIC OPTIMIZATION PROBLEMS,
I N PARTICULAR FOR
STOCHASTIC PROGRAMS WITH RECOURSE
John B i r g e and Roger J-B. Wets
I n d u s t r i a l and O p e r a t i o n s E n g i n e e r i n g Department o f Mathematics U n i v e r s i t y o f M i c h i g a n U n i v e r s i t y o f Kentucky
ABSTRACT
Various a p p r o x i m a t i o n schemes f o r s t o c h a s t i c o p t i m i z a t i o n problems i n v o l v i n g e i t h e r approximates o f t h e p r o b a b i l i t y mea- sures and/or approximates o f t h e o b j e c t i v e f u n c t i o n a l , a r e i n - v e s t i g a t e d . We d i s c u s s t h e i r p o t e n t i a l imp1 e m e n t a t i o n as p a r t o f g e n e r a l procedures f o r s o l v i n g s t o c h a s t i c programs w i t h r e - course.
Supported i n p a r t by g r a n t s of t h e N a t i o n a l Science Foundation.
1.
INTRODUCTION
We t a k e
( 1 . 1 ) f i n d x E Rn t h a t m i n i m i z e s E f ( x ) = E{f ( x ,
6
( o ) ) )a s p r o t o t y p e f o r t h e c l a s s o f s t o c h a s t i c o p t i m i z a t i o n problems u n d e r i n v e s t i g a - t i o n , where
N N
5
i s a random v e c t o r which maps t h e p r o b a b i l i t y s p a c e (R,A,P) on (R, B
, F ) w i t h F t h e d i s t r i b u t i o n f u n c t i o n and c RN t h e s u p p o r t o f t h e p r o b a b i l i t y m e a s u r e i n d u c e d by5
( i . e . F i s t h e s e t o f p o s s i b l e v a l u e s assumed b y 5 ) , andf : R " X R ~ + R u { + a ) i s a n e x t e n d e d r e a l - v a l u e d f u n c t i o n . We s h a l l assume:
( 1 . 3 ) f o r a l l x , WH f ( x , C ( o ) ) i s m e a s u r a b l e
,
and t h e f o l l o w i n g i n t e g r a b i l i t y c o n d i t i o n :
( 1 . 4 ) i f P [ o
I
f ( x , C ( w ) ) < + a ] = l t h e n E f ( x ) < + a.
We r e f e r t o E f = ~ { f ( 0
,c
( a ) ) } a s a n e x p e c t a t i o n f u n c t i o n a l . Note t h a t i t c a n a l s o b e e x p r e s s e d a s a L e b e s g u e - S t i e l t j e s i n t e g r a l w i t h r e s p e c t t o F:r
A wide v a r i e t y o f s t o c h a s t i c o p t i m i z a t i o n p r o b l e m s f i t i n t o t h i s ( a b s t r a c t ) framework ; i n p a r t i c u l a r s t o c h a s t i c programs w i t h ( f i x e d ) recourse [ I.]
( 1 . 6 ) f i n d x r R:' s u c h t h a t Ax = b
,
and z = c x + O_(x) i s m i n i m i z e d where A i s an m xn - m a t r i x , b E R m l ,
1 1
a n d , t h e recourse f u n c t i o n i s d e f i n e d by
The ( m 2 x n 2 ) - m a t r i x W i s c a l l e d t h e recourse matrix. F o r e a c h w : T(w) i s m xn 2 1' q(w) E Rn2 and h(w) E R m 2 . P i e c i n g t o g e t h e r t h e s t o c h a s t i c components o f t h e p r o - blem, we g e t a v e c t o r
<
€ RN w i t h N = n + m + (m r n ) , and2 2 2 1
We s e t
( 1 . 9 ) f ( x , < ) = r c x + Q ( x , < ) i f AX = b , x 1 0 ,
1 + -
o t h e r w i s e.
P r o v i d e d t h e r e c o u r s e problem i s a . s . bounded, i . e . ( 1 . 1 0 ) P [ w 1 3 ~ s u c h t h a t ~ W s q ( ~ ) ] = I ,
which we assume h e n c e f o r t h , t h e f u n c t i o n Q and t h u s a l s o f , d o e s n o t t a k e on t h e v a l u e - m
.
The m e a s u r a b i l i t y o f f ( x , * ) f o l l o w - d i r e c t l y from t h a t o f6
t+ Q ( x , < )[ l , S e c t i o n 31. I f
6
h a s f i n i t e s e c o n d moments, t h e n Q(x) i s f i n i t e whenever w - Q ( x , < ( w ) ) i s f i n i t e [ l , Theorem 4.11 and t h i s g u a r a n t e e s c o n d i t i o n ( 1 . 4 ) .Much i s known a b o u t problems o f t h i s t y p e [ l ] . The p r o p e r t i e s of f a s d e - f i n e d t h r o u g h ( 1 . 9 ) , q u i t e o f t e n m o t i v a t e and j u s t i f y t h e c o n d i t i o n s u n d e r which we o b t a i n v a r i o u s r e s u l t s . The r e l e v a n t p r o p e r t i e s a r e
( 1 . 1 1 ) ( h , T ) t + Q ( x , < = ( q , h , T ) ) i s a p i e c e w i s e l i n e a r convex f u n c t i o n f o r a l l f e a s i b l e x c K = K n K 2 ,
1 where
K 1 = { X
1
AX=^, X L O )K 2 = { X IY<(w) E = ,
-
3 y ~ O s u c h t h a t W y = h ( w ) - ~ ( w ) x ) ,( 1 . 1 2 ) q t + Q ( x , < = ( q , h , T ) ) i s a concave p i e c e w i s e l i n e a r f u n c t i o n
,
and
( 1 . 1 3 ) x + Q ( x , < ) i s a convex p i e c e w i s e l i n e a r f u n c t i o n which i m p l i e s t h a t
(1.14) x!+O_(x) is a Lipschitzian convex function
,
finite on K as follows from the integrability condition on c(-).
2 '
When T is nonstochastic, or equivalently does not depend on w , it is some- times useful to work with a variant formulation of (1.6). With T = T(w) for all w, we obtain
(1.15) find x E ,':R
x
E Rm2 such thatAX = b
,
TX =x ,
and= cx + y
(x)
is minimized where(1.16) ~ ( X ) = E { $ ( X , E ( U ) ) } = ~ $ ( X ~ E ( ~ ) ) P ( ~ ~ ) and
This formulation stresses the fact that choosing x corresponds to generating a tender X = Tx to be "bid" against the outcomes h(w) of random events. The func- tions $ and Y have basically the same properties as
Q
and C?, replacing naturally the set K2 by the set L2 = {x-Tx1
x E K21
={X I
Y h (a) e-
zh, 3 Y 1 0 such that WY = h(w) -Tx}The function f is now given by
L +-
otherwise.
A significant number of applications have the function $ separable, i.e.
$(X.S) = qi(Xi,Ei) such as in stochastic programs with simple recourse [l, Section 61. This will substantially simplify the implementation of various approximation schemes described below. When separability is not at hand, it will sometimes be useful to introduce it, by constructing appropriate approximates for
$ or
Q,
see Section 3.Another common f e a t u r e o f s t o c h a s t i c o p t i m i z a t i o n problems, t h a t one should n o t l o s e t r a c k o f when d e s i g n i n g approximation schemes, i s t h a t t h e random b e h a v i o r o f t h e s t o c h a s t i c elements o f t h e problem can o f t e n be t r a c e d back t o a few i n d e - pendent random v a r i a b l e s . T y p i c a l l y
(1.19) 5 ( ~ ) = C ~ ( ~ ) ~ ~ + C ~ ( ~ ) E ~ + ~ * * + S ~ ( W ) ~ M where t h e
a r e independent r e a l - v a l u e d random v a r i a b l e s , and
a r e f i x e d v e c t o r s . In f a c t many a p p l i c a t i o n s
- -
such a s t h o s e i n v o l v i n g s c e n a r i o a n a l y s i s- -
i n v o l v e j u s t one such random v a r i a b l e < ( * ) ; n a t u r a l l y , t h i s makes t h e components o f t h e random v e c t o r 5 ( * ) h i g h l y dependent. L a s t , b u t n o t l e a s t , o n l y r a r e l y do we have i n p r a c t i c e a d e q u a t e s t a t i s t i c s t o model w i t h s u f f i c i e n t a c c u r - acy j o i n t phenomena i n v o l v i n g i n t r i c a t e r e l a t i o n s h i p s between t h e components of5.
Hence, we s h a l l d e v o t e most o f o u r a t t e n t i o n t o t h e independent c a s e , remaining a t a l l t i m e s v e r y much aware of t h e c o n s t r u c t i o n ( 1 . 1 9 ) .
T h i s w i l l s e r v e a s background t o o u r s t u d y o f approximation schemes f o r c a l c u l a t i n g
A f t e r t a k i n g c a r e o f some g e n e r a l convergence r e s u l t s ( S e c t i o n 2 ) , we b e g i n o u r s t u d y w i t h a d e s c r i p t i o n o f p o s s i b l e approximates o f f i n t h e c o n t e x t o f s t o c h a s - t i c programs w i t h r e c o u r s e ( S e c t i o n 3 . ) We t h e n examine t h e p o s s i b i l i t y of ob- t a i n i n g lower o r upper bounds on E f by means o f d i s c r e t i z a t i o n ( o f t h e p r o b a b i l i t y measure) u s i n g c o n d i t i o n a l e x p e c t a t i o n s ( S e c t i o n 41, measures w i t h e x t r e m a l
s u p p o r t ( S e c t i o n 5)
,
e x t r e m a l m e a s u r e s ( S e c t i o n 6) o r m a j o r i z i n g p r o b a b i l i t y mea- s u r e s ( S e c t i o n 7 ) . I n e a c h c a s e we a l s o s k e t c h o u t t h e i m p l e m e n t a t i o n o f t h e r e - s u l t s i n t h e framework o f s t o c h a s t i c programs w i t h r e c o u r s e , r e l y i n g i n some c a s e s on t h e a p p r o x i m a t e s t o f o b t a i n e d i n S e c t i o n 3 . I n t h e l a s t s e c t i o n , we g i v e some f u r t h e r e r r o r bounds f o r i n f E f t h a t r e q u i r e t h e a c t u a l c a l c u l a t i o n o f E ( x ) a t some p o i n t s .f
The purpose o f t h i s s e c t i o n i s t o f r e e u s a t once from any f u r t h e r d e t a i l e d argumentation i n v o l v i n g convergence o f s o l u t i o n s , i n f i m a , and s o on. To do s o we r e l y on t h e t o o l s p r o v i d e d by e p i - c o n v e r g e n c e . Let {g; g v
,
v = l , ..
. ) be a c o l - l e c t i o n of f u n c t i o n s d e f i n e d on R n w i t h v a l u e s i nR =
[-ca,+w].
The sequencev n
{g
,
v = l ,. . . I
i s s a i d t o epi-converge t o g i f f o r a l l X E R,
we have(2 1 ) v v v
l i m i n f g (x ) 2 g ( x ) f o r a l l {x
,
v = l ,. . . I
converging t o x,
Vt"
and
(2.2) v v v
t h e r e e x i s t s {x
,
v = l ,. .
. ) converging t o x such t h a t l i m s u p g (x ) 2 g ( x ).
Vt03
Note t h a t any one o f t h e s e c o n d i t i o n s i m p l i e s t h a t g, t h e epi-limit o f t h e g v
,
i s n e c e s s a r i l y lower s e m i c o n t i n u o u s . The name epi-convergence comes
from t h e f a c t t h a t t h e f u n c t i o n s {g v
,
v = l ,. .
.) e p i - c o n v e r g e t o g i f and o n l y i f t h e s e t s { e p i g v,
v = l ,. . . 1
converge t o e p i g = { ( x , a )I
g ( x ) 5 a); f o r more d e t a i l s c o n s u l t [ 2 , 3 ] . Our i n t e r e s t i n epi-convergence stems from t h e f o l l o w i n g p r o p e r - t i e s [ 4 ] .2.3 THEOREM. Suppose a sequence of functions v = l
, . . .
) epi-converges to g.Then
( 2 . 4 ) v
l i m s u p ( i n f g ) 2 i n f g
,
Vt"
and, if
k v
x E argmin g Vk = {x
I
g x 2 i n f g V k )fop some subsequence of functions {gVk, k = l , .
.
. ) and x = l i m x k,
it fottowstkzt
k-x E argmin g
,
and l i m ( i n f g vk ) = i n f g
.
k-
Moreover, if argmin g t 8, then l i m ( i n f g') = i n f g if and only if x c argmin g
Vt"
i m p l i e s the e x i s t e n c e o f sequences {E" 0, v = l , .
. .
) and { x V,
v = l ,. .
. ) w i t hl i m ~ = 0 , und l i m x = x , v
Vt03 V v-
such t h a t for a l l v = l ,
...,
V V
x E E -argmin g = {x
I
g ( x ) I i n f g V + ~ v ).
V V
2 . 5 COROLLARY. Suppose a sequence o f functions {g V
,
v = l ,. . . I
epi-converges t o g,and
there e x i s t s a bounded s e t D such t h a targmin g V n D t 0
f o r a l l v s u f f i c i e n t l y large. Then l i m ( i n f g ) = i n f g
vtco V
and t h e minimwn o f g i s a t t a i n e d a t some point i n t h e closure o f D:
PROOF. S i n c e D i s bounded, i t f o l l o w s t h a t t h e r e e x i s t s a bounded s e q u e n c e {x V
,
v = l ,. . .
) w i t hv v
x E argmin g n D
.
T h i s means t h a t a s u b s e q u e n c e c o n v e r g e s { x v k , 1 , .
.
t o a p o i n t x b o t h i n t h e c l o s u r e o f D and i n argmin g a s f o l l o w s from e p i - c o n v e r g e n c e . Theorem 2 . 3 a l s o y i e l d sl i m g V k ( x Vk ) = g ( x ) = i n f g
.
k-
T h e r e r e m a i n s o n l y t o a r g u e t h a t t h e e n t i r e s e q u e n c e { ( i n f g V ) , v = l , . . . ) c o n v e r g e s t o i n f g . But t h i s s i m p l y f o l l o w s from t h e o b s e r v a t i o n t h a t t h e p r e c e d i n g a r g u - ment a p p l i e d t o any s u b s e q u e n c e y i e l d s a f u r t h e r s u b s e q u e n c e c o n v e r g i n g t o i n f g . I]
The f o l l o w i n g p r o p o s i t i o n p r o v i d e s v e r y u s e f u l c r i t e r i a f o r v e r i f y i n g e p i - c o n v e r g e n c e .
v -
2 . 6 PROPOSITION. [S, P r o p o s i t i o n 3.121 Suppose {g : R n + R , v = l ,
... 1
i s a col- l e c t i o n o f functions pointwise converging t o g,i .
e . for a l l x, g ( x ) = l i m gV ( x ).
v-
Then t h e gv epi-converge t o g, i f they are monotone increasing, or monotone de- creasing w i t h g lower semicontinuous i n t h i s l a t t e r case.
For e x p e c t a t i o n f u n c t i o n a l s , we o b t a i n t h e n e x t a s s e r t i o n as a d i r e c t c o n s e - quence o f t h e d e f i n i t i o n o f e p i - c o n v e r g e n c e and F a t o u ' s lemma.
2 . 7 THEOREM. Suppose { f ; f v , v = l , .
. .
} i s a c o l l e c t i o n o f functions defined on RnxQ w i t h values i n R u { + a ) s a t i s f y i n g conditions ( 1 . 3 ) and ( 1 . 4 ) , such t h a t for a l l5
EE
t h e sequence { f V ( ,<),
v = l ,. . .
) epi-converges t o f ( - , c ).
Suppose more- over t h a t t h e functions f V are bounded below uniformly. Then t h e expectation functionals E epi-converge t o Ef V f '
When i n s t e a d o f a p p r o x i m a t i n g t h e f u n c t i o n a l f , we a p p r o x i m a t e t h e p r o b a - b i l i t y measure P, we g e t t h e f o l l o w i n g g e n e r a l r e s u l t t h a t s u i t s o u r n e e d s i n most a p p l i c a t i o n s , s e e [ 6 , Theorem 3.91
,
[ 7 , Theorem 3.31.
2.8 THEOREM. Suppose {P v = l ,
. . . I
i s a sequence of p r o b a b i l i t y measures con- v'
verging i n d i s t r i b u t i o n t o t h e p r o b a b i l i t y measure P d e f i n e d on Q, a separable m e t r i c space w i t h
A
t h e Bore2 sigma-field. Letbe continuous i n w for each fixed x i n K , where
and l o c a l l y L i p s c h i t z i n x on K w i t h L i p s c h i t z constant independent o f w. Sup- pose moreover t h a t for any X E K and E > O t h e r e e x i s t s a compact s e t SE and v such
E
t h a t for a l l v 2 v
E
R\SE
mul with V={w
(
f ( x , w ) =+a), P(V) > O if and only if PV(V) '0. Then the sequence of expectation fvnctionats {E:, v = l ,. . .
) epi- and poinhjise converges to E f J wherePROOF. We b e g i n by showing t h a t t h e E V p o i n t w i s e converge t o E F i r s t l e t
f f '
x E K and s e t
From ( 2 . 9 ) ' i t f o l l o w s t h a t f o r a l l E > O , t h e r e i s a compact s e t SE and i n d e x vE s u c h t h a t f o r a l l v r vE
[
Ig(w) IPv(dw) < E.
Let ME
- -
SupwEsE I g ( w ) ( . We know t h a t M i s f i n i t e s i n c e S i s compact and g i sE E
c o n t i n u o u s , r e c a l l t h a t X E K . Let g be E a t r u n c a t i o n o f g d e f i n e d by
The f u n c t i o n g E i s bounded and c o n t i n u o u s and we have t h a t f o r . a l 1 we R
Hence from t h e convergence i n d i s t r i b u t i o n o f t h e Pv
r 7
and a l s o f o r a l l v l v
E '
R\SE Now l e t
We have t h a t f o r a l l
v
2 vEand a l s o t h a t
Combining t h e two l a s t i n e q u a l i t i e s w i t h (2.10) shows t h a t f o r a l l E > O , t h e r e e x i s t s
v
such t h a t f o r a1 1v
2v
E E
and t h u s f o r a 1 1 x E K ,
l i m E
v
( x ) = E f ( x ).
Vt03 f
I f x
4
K, t h i s means t h a twhich a l s o means t h a t f o r a l l
v
from which i t f o l l o w s t h a t f o r a l l
v
l i m E
v
(x) = +a = E f ( x ).
V*W f
n
v
And t h u s , f o r a l l X E R
,
E f ( x ) = l i m E f ( x ) . T h i s g i v e s us n o t o n l y p o i n t w i s ev-
convergence, b u t a l s o c o n d i t i o n ( 2 . 2 ) f o r e p i - c o n v e r g e n c e .
To c o m p l e t e t h e p r o o f , i t t h u s s u f f i c e s t o show t h a t c o n d i t i o n ( 2 . 1 ) i s s a t i s f i e d f o r a l l X E K . The f u n c t i o n x w f ( x , < ( w ) ) i s L i p s c h i t z i a n on K , w i t h L i p s c h i t z c o n s t a n t L i n d e p e n d e n t o f w . F o r any p a i r x , x w i n K , we h a v e t h a t f o r a l l w
which i m p l i e s t h a t
Let u s now t a k e xw as p a r t o f a s e q u e n c e {x w
,
w = l , ..
.) c o n v e r g i n g t o x . I n t e g r a t - i n g on b o t h s i d e s o f t h e p r e c e d i n g i n e q u a l i t y and t a k i n g lirn i n f , we g e tLH03
w w
E ( x ) = l i m E f ( x ) - L l i m d i s t ( x , x )
f w- LH03
= l i m i n f (E;(X) - L d i s t ( x , x w ) )
LH03
w V
lirn i n f E f ( x )
,
LH03
which c o m p l e t e s t h e p r o o f .
O
2.11 APPLICATION. Suppose {P w = l ,
...)
i s a sequence of probability measures w'
t h a t converge i n d i s t r i b u t i o n t o P, a l l w i t h compact support $2. Suppose
w i t h the recourse function Q defined by ( 1 . 8 ) and Q by ( 1 . 7 ) . Then the Qw both e p i - and pointwise converge t o Q.
I t s u f f i c e s t o o b s e r v e t h a t t h e c o n d i t i o n s o f Theorem 2 . 8 a r e s a t i s f i e d . The c o n t i n u i t y o f Q(x,S) w i t h r e s p e c t t o
<
( f o r x E K ) f o l l o w s from ( 1 . 1 1 ) and2
( 1 . 1 2 ) . The L i p s c h i t z p r o p e r t y w i t h r e s p e c t t o x i s o b t a i n e d from [ l , Theorem 7 . 7 1 ; t h e p r o o f o f t h a t Theorem a l s o shows t h a t t h e L i p s c h i t z c o n s t a n t i s i n d e p e n d e n t o f
5,
c o n s u l t a l s o [ 8 ] .2.12 IMPLEMEIVTATION. From t h e p r e c e d i n g r e s u l t s i t f o l l o w s t h a t we have been g i v e n g r e a t l a t i t u d e i n t h e c h o i c e o f t h e p r o b a b i l i t y measures t h a t approximate P . However, i n what f o l l o w s we c o n c e r n o u r s e l v e s a l m o s t e x c l u s i v e l y w i t h d i s c r e t e p r o b a b i l i t y measures. The b a s i c r e a s o n f o r t h i s i s t h a t t h e form o f f ( x , c ) - - o r Q(x,c) i n t h e c o n t e x t o f s t o c h a s t i c programs w i t h r e c o u r s e - - r e n d e r s t h e numer- i c a l e v a l u a t i o n o f E f ( o r Ef) p o s s i b l e o n l y i f t h e i n t e g r a l i s a c t u a l l y a ( f i n i t e ) v sum. Only i n h i g h l y s t r u c t u r e d problems, s u c h a s f o r s t o c h a s t i c programs w i t h s i m p l e r e c o u r s e [ 9 ] , may i t b e p o s s i b l e and p r o f i t a b l e t o u s e o t h e r a p p r o x i m a t i n g measures.
3. APPROXIMATING T H E RECOURSE FUNCTION Q
When f i s convex i n
5,
i t i s p o s s i b l e t o e x p l o i t t h i s p r o p e r t y t o o b t a i n s i m p l e b u t v e r y u s e f u l l o w e r bounding a p p r o x i m a t e s f o r Ef '
3 . 1 PROPOSITION.
Suppose <*
f ( x , 5 )i s convex,
( 5 R,
R = l ,. . .
, v )i s a f i n i t e c o l l e c - t i o n of points i n , and for
R = 1 ,. . . , ,
i . e .
v'i s a subgradient o f
f ( x , - )a t 5'. Then
( 3 . 2 ) E f (XI 2 E{ max [vR<(w) + ( f ( x , t R )
-
v k R )I} .
1 r Rrv
PROOF. To s a y t h a t v R i s a s u b g r a d i e n t o f t h e convex f u n c t i o n o f f ( x , 0) a t
5
G,
means t h a t
S i n c e t h i s h o l d s f o r e v e r y R, we o b t a i n
R R R R
f ( x , S ) 2 max [ v
5 +
( f ( x , S - v5 11 .
1 l Rlv
I n t e g r a t i n g on b o t h s i d e s y i e l d s ( 3 . 2 ) .
O
3.3 APPLICATICIN.
Consider t h e s t o c h a s t i c program w i t h recourse
( 1 . 6 )and suppose
R R E
t h a t only
hand
Tare s t o c h a s t i c . Let
( 5 = (h ,T ),
R = l ,. . .
, v )be a f i n i t e number of r e a l i z a t i o n s o f
hand
T, x e K 2and f o r
k = l ,...,
v,R R R
n E argmax [ n ( h - T x)
1
nW I q ].
( 3 . 4 ) R
e ( x ) 2 E{ max n (h (o) - T(w) x) l l R l v
T h i s i s a d i r e c t c o r o l l a r y o f P r o p o s i t i o n 3 . 1 . We g i v e a n a l t e r n a t i v e p r o o f which c o u l d be o f some h e l p i n t h e d e s i g n o f t h e i m p l e m e n t a t i o n . S i n c e x e K
2 '
f o r e v e r y c = ( h , T ) i n 3, t h e l i n e a r program ( 3 . 5 ) f i n d n e R~~ s u c h t h a t nW 5 q
and w = n ( h - Tx) i s maximized
i s bounded, g i v e n n a t u r a l l y t h a t i t i s f e a s i b l e a s f o l l o w s from a s s u m p t i o n ( 1 . 1 0 ) . Hence f o r R = l ,
. . . ,
v ,R R R R
Q ( x , S
1
= n (h - T x ),
and moreover s i n c e n R i s a f e a s i b l e s o l u t i o n o f t h e l i n e a r program ( 3 . 5 ) , f o r a l l
C E
3Q ( x , c ) 2 n (h R - Tx)
.
S i n c e t h i s h o l d s f o r e v e r y R , we g e t Q ( x , c ) 2 max n (h - T x ) R
.
l<R<v'
I n t e g r a t i n g on b o t h s i d e s y i e l d s ( 3 . 4 ) .
3.6
IMPLEMENTATION.
I n g e n e r a l f i n d i n g i n e x p r e s s i o n ( 3 . 4 ) , t h e maximum f o r e a c h5 --
o r e q u i v a l e n t l y f o r e a c h ( h , T ) e 3- -
c o u l d b e much t o o i n v o l v e d . But we may a s s i g n t o e a c h re a s u b r e g i o n of f, w i t h o u t r e s o r t i n g t o ( e x a c t ) maximiza- t i o n . The l o w e r bound may t h e n n o t b e a s t i g h t a s ( 3 . 4 ) , b u t we c a n r e f i n e i t by t a k i n g s u c c e s s i v e l y f i n e r and f i n e r p a r t i t i o n s . However, o n e s h o u l d n o t f o r g e t t h a t ( 3 . 4 ) i n v o l v e s a r a t h e r s i m p l e i n t e g r a l and t h e e x p r e s s i o n t o t h e r i g h t c o u l d b e e v a l u a t e d n u m e r i c a l l y t o a n a c c e p t a b l e d e g r e e o f a c c u r a c y , w i t h o u t m a j o r d i f - f i c u l t i e s . Note t h a t t h e c a l c u l a t i o n o f t h i s l o w e r bound i m p o s e s no l i m i t a t i o n s on t h e c h o i c e o f t h e5
R.
However, i t i s o b v i o u s t h a t a w e l l - c h o s e n s p r e a d o f t h eR R
( 5
,
R = l ,...,
v ) w i l l y i e l d a b e t t e r a p p r o x i m a t i o n . F o r example, t h e5
c o u l d b et h e c o n d i t i o n a l e x p e c t a t i o n o f c ( - ) w i t h r e s p e c t t o a p a r t i t i o n i n g S = {S R = l ,
...,
v ) R 'o f
:
which a s s i g n s t o each S a b o u t t h e same p r o b a b i l i t y . Also t h e u s e o f a Rl a r g e r c o l l e c t i o n o f p o i n t s , i . e . i n c r e a s i n g v , w i l l a l s o y i e l d a b e t t e r lower bound.
3.7 CONVERGENCE. Suppose that c w f ( x , S ) is convex. For each v = l ,
...,
ZetS V = {Se, v = l , ,} denote a partition of t with v
the conditionaZ expectation of 5 ( * ) given
sV
Suppose moreover that S v p + 1 R'and that (3.8)
Then, with v V R e 8 f ( x , e V " and
5
( 3 . 9 ) v max [ v R V S ( u ) + f ( x , E V R )
-
v l r R l Lv
we have that the sequence of functions {E:, v = l , .
.
. } is monotone increasing, and for aZZ x:Ef(x) = l i m Ef(x) v
.
v-
Hence the sequence {EV v = l ,
. .
. } is both pointwise- and epi-convergent.f '
PROOF. From P r o p o s i t i o n 3.1, i t f o l l o w s t h a t EVf r E f f o r a l l v . The i n e q u a l i t y
t h e n f o l l o w s s i m p l y from t h e f a c t t h a t S = S v . Now o b s e r v e t h a t
(3.10) vR vR v
max [vvRS + f ( x , ~ ~ ' ) - v
5 1
2 g ( x , ~ ) lrRrLv
where gv i s d e f i n e d a s f o l l o w s :
It follows that
L
which gives us
v v
Ef(x) 2 lim E (x) 2 lim E{g (x,S(w))) = Ef(x) ;
LH03 f
w
the last equality following from assumption (3.8).
We have thus shown that the sequence { E ~ , v v = l , is monotone increasing and pointwise converges, and this implies epi-convergence, see Proposition 2.6.
O
If f(x,=) is concave, the inequality in (3.2) is reversed and, instead of a lower bound on Ef, we obtain an upper bound. In fact, we can again use Proposi- tion 3.1, but this time applied to -f.
3.11 APPLICATION. Consider t h e s t o c h a s t i c program w i t h recourse (1.6) and sup-
R R
pose t h a t only t h e v e c t o r q i s s t o c h a s t i c . Let
15
= q,
R=l,...,
v) be a f i n i t e nwnber o f r e a l i z a t i o n s o f q, x E K2 and f o ra=
1,. . . ,
v(3.12) Q(x) 5 E{ min q (w) ye)
.
1 5 R5v
Again this is really a corollary of Proposition 3.1. A slightly different proof proceeds as follows: Note that for all q =
S
EE,
for every R, is a feasible, but not necessarily optimal, solution of the linear programfind y E R : ~ such that Wy = p-Tx, and w = q y is minimized.
Hence
Q(x,S> I min q~
R
llR5vfrom which (3.12) follows by integrating on both sides.
The r e m a r k s made a b o u t I m p l e m e n t a t i o n 3 . 6 a n d t h e a r g u m e n t s u s e d i n C o n v e r g e n c e 3 . 7 s t i l l a p p l y t o t h e c o n c a v e c a s e s i n c e we a r e i n t h e same s e t t i n g a s b e f o r e p r o v i d e d we work w i t h - f o r -Q.
P r o p o s i t i o n 3 . 1 p r o v i d e s u s w i t h a l o w e r bound f o r Ef when < * f ( x , < ) i s convex, t h e n e x t r e s u l t y i e l d s a n u p p e r bound.
3.13 PROPOSITION.
Suppose<
t+ f ( x , <) i s convex, { < R,
R = l ,. . . ,
v ) i s a f i n i t e c o l - l e c t i o n of points i n:.
Thenwhere
-
v( 3 . 1 5 ) f ( x , < ) v = i n f
I f t h e function < B f ( x , < ) i s sublinear, the f v can be defined a s follows:
v v
( 3 . 1 6 ) f ( x , < ) v = i n f
L
(Note t h a t f V ( x , <) i s +m i f t h e corresponding program i s i n f e a s i b l e . )
PROOF.
C o n v e x i t y i m p l i e s t h a t f o r a l l h 1- > O ,. . . ,
hvrO w i t h v"
= 1, and<
= he<' we h a v efrom which ( 3 . 1 4 ) f o l l o w s u s i n g ( 3 . 1 5 ) . S u b l i n e a r i t y ( c o n v e x i t y and p o s i t i v e homogeneity) a l s o y i e l d s ( 3 . 1 7 ) b u t t h i s t i m e w i t h o u t
1'
h = 1, and t h i s i nR = l R t u r n y i e l d s ( 3 . 1 4 ) u s i n g ( 3 . 1 6 ) t h i s t i m e .
O
3.18 APPLICATION.
Ray functions. C o n s i d e r t h e s t o c h a s t i c program w i t h r e c o u r s e i n t h e form ( 1 . 1 5 ) and s u p p o s e t h a t o n l y h i s s t o c h a s t i c , i . e . w i t h f i x e d m a t r i x T a n d r e c o u r s e c o s t c o e f f i c i e n t s q . Now suppose t h a t f o r g i v e n X , we h a v e t h e v a l u e s o fR R
{ $ J ( x , < =h ) , R = l , .
.
. , v ) f o r a f i n i t e c o l l e c t i o n o f r e a l i z a t i o n s o f h ( * ) .Let
5
E I and d e f i n e ( 3 . 1 9 ) $'(x, S) = i n fR= 1 Then
Y(x) 5 y V ( x ) = $ v ( x , S ( w ) ) p ( d w )
The above f o l l o w s from t h e s e c o n d p a r t o f P r o p o s i t i o n 3 . 1 3 p r o v i d e d we o b s e r v e t h a t from t h e d e f i n i t i o n ( 1 . 1 7 ) o f $, we h a v e t h a t
i s s u b l i n e a r . From t h i s i t f o l l o w s t h a t f o r any A E R:
whenever
and t h i s l e a d s t o t h e c o n s t r u c t i o n of i n ( 3 . 1 9 ) . I]
3.20 IMPLEMENTATION. F i n d i n g f o r e a c h
5,
t h e o p t i m a l v a l u e o f a l i n e a r program a s r e q u i r e d by t h e d e f i n i t i o n o f i n ( 3 . 1 9 ) , c o u l d i n v o l v e much more work t h a n i s a p p r o p r i a t e t o i n v e s t i n t h e c o m p u t a t i o n o f a n u p p e r bound. One way t o remedy t h i s i s t o s u b d i v i d eE
s u c h t h a t e a c h5
i s a u t o m a t i c a l l y a s s i g n e d t o a p a r t i c u l a r r e g i o n spanned by a s u b s e t o f t h e ( 5 R,
R = l ,...,
v ) o r t o t h e s u b s e t whose p o i n t s a r e s u c h t h a tOne c a s e i n which a l l o f t h i s f a l l s n i c e l y i n t o p l a c e i s when a s t o c h a s t i c program w i t h r e c o u r s e o f t y p e ( 1 . 1 5 ) c a n b e a p p r o x i m a t e d by a s t o c h a s t i c program w i t h simple recourse [ l , S e c t i o n 61 where t h e f u n c t i o n $ ( x , C ) i s s e p a r a b l e ,
and
+ -
h e r e < . = ( q . , q . , h . ) . The f u n c t i o n
Y
i s t h e n a l s o s e p a r a b l e and c a n be e x p r e s s e d a s1 1 1 1
m 2
where
( T h i s i s t h e
linear
v e r s i o n o f t h e s i m p l e r e c o u r s e p r o b l e m . )3.23
APPLICATION. Approximation
bysimple recourse.
C o n s i d e r a s t o c h a s t i c program w i t h r e c o u r s e o f t h e t y p e ( 1 . 1 5 ) , w i t h o n l y h s t o c h a s t i c and c o m p l e t e r e c o u r s e[ l , S e c t i o n 61, t h i s means t h a t t h e r e c o u r s e m a t r i x W i s s u c h t h a t p o s W = { t
1
t = W y , y 2 O ) = R~~,
i . e . t h e r e c o u r s e problem i s f e a s i b l e w h a t e v e r be h o r X. F o r i = l ,
...,
m 2 , d e f i n e ( 3 . 2 4 )ql
= i n f { q yI
Wy = e i,
y 2 0 ),
and
( 3 . 2 5 ) q I = i n f { q y
1
W = - e i,
y 2 0 ),
where e i s t h e u n i t v e c t o r w i t h a i 1 i n t h e i t h p o s i t i o n , i . e .
i T
e = (0
,...,
O , l , O,...,
0).
The r e c o u r s e f u n c t i o n $ ( x , S ) i s t h e n a p p r o x i m a t e d by t h e r e c o u r s e f u n c t i o n ( 3 . 2 1 ) o f a s t o c h a s t i c program w i t h s i m p l e r e c o u r s e u s i n g f o r q and q i t h e +
i
v a l u e s d e f i n e d by ( 3 . 2 4 ) a n d ( 3 . 2 5 ) . T h i s i s a s p e c i a l c a s e o f t h e r a y f u n c - t i o n s b u i l t i n A p p l i c a t i o n 3 . 1 8 ; e a c h (S-X) f a l l s i n a g i v e n o r t h a n t and i s t h u s a u t o m a t i c a l l y a s s i g n e d a p a r t i c u l a r p o s i t i v e l i n e a r c o m b i n a t i o n o f t h e c h o s e n
p o i n t s ( t e i
-x,
i = l,...,
m2). To improve t h e approximation we have t o i n t r o d u c e a d d i t i o n a l v e c t o r s5
R,
which b r i n g s us back t o t h e more g e n e r a l s i t u a t i o n de- s c r i b e d i n A p p l i c a t i o n 3.18.3.26 A P P L I C A T I O N . Consider a s t o c h a s t i c program w i t h r e c o u r s e o f t y p e ( 1 . 1 5 ) , w i t h o n l y q s t o c h a s t i c . The f u n c t i o n
i s n o t o n l y concave and p o l y h e d r a l ( 1 . 1 2 ) , i t i s a l s o p o s i t i v e l y homogeneous.
R R
For any f i n i t e c o l l e c t i o n
15
= q,
R = l ,. . .
, v ) we have t h a tR R R
T h i s a g a i n f o l l o w s d i r e c t l y from P r o p o s i t i o n 3.13; n o t e t h a t @ ( x , q ) = q y where
R R
y E argmin[q y
I
Wy = h-X, y 2 01.3.28 IMPLEMENTATIOIV. C a l c u l a t i n g f o r each q, t h e upper bound p r o v i d e d by (3.27) may be p r o h i b i t i v e . We c o u l d a s s i g n each q ~
E
t o some s u b r e g i o n o f 5 spanned by t h e p o s i t i v e combinations o f some o f t h e {q R,
R = l ,...,
v ) . Such a bound i s much e a s i e r t o o b t a i n b u t c l e a r l y n o t a s s h a r p a s t h a t g e n e r a t e d by ( 3 . 2 7 ) .Another approach t o g e t t i n g upper and lower bounds f o r s t o c h a s t i c programs w i t h r e c o u r s e i s t o r e l y on t h e
p a i r s programs
a s i n t r o d u c e d i n [ l o , S e c t i o n 41.One r e l i e s a g a i n on c o n v e x i t y p r o p e r t i e s and once a g a i n one needs t o d i s t i n g u i s h between (h,T) s t o c h a s t i c , and q s t o c h a s t i c . To b e g i n w i t h , l e t u s c o n s i d e r h , and
A A A
T s t o c h a s t i c . For e v e r y (h,T) =
5
E E , and (h,T) =5
E c o Z ( t h e convex h u l l of ),
l e t (3.29) p ( i , 5 ) = i n f [ c x + ;q? + (l-G)qy5]such t h a t Ax = b
w i t h
6
E [ 0 , 1 ] . I f ( 1 . 6 ) i s s o l v a b l e , s o i s (3.29) a s f o l l o w s from [ l , Theorem 4 . 6 1 . Suppose x s o l v e s ( 1 . 6 ) and f o r a l l ( = ( h , T ) , l e t 0Y 0
(6)
E argminyERn2 [qy1
Wy = h-Tx].
+
I t i s well-known t h a t can be chosen s o t h a t < b y 0 ( < ) i s measurable [ l , S e c t i o n 31. Now suppose
- - -
5
= (h,T) = E{S},
and
-
Y = E{r0(5>
1 .
0 - 0
The t r i p l e (x , y , y ( 5 ) ) i s a f e a s i b l e , b u t n o t n e c e s s a r i l y o p t i m a l , s o l u t i o n o f
A A - -
t h e l i n e a r program (3.29) when (h ,T) = (h ,T)
.
Henceand i n t e g r a t i n g on b o t h s i d e s , we o b t a i n
T h i s bound can be r e f i n e d i n many ways: f i r s t , i n s t e a d o f j u s t u s i n g one p o i n t
5,
one could u s e a c o l l e c t i o n of p o i n t s o b t a i n e d a s c o n d i t i o n a l e x p e c t a t i o n s o f a p a r t i t i o n of E , and c r e a t e a p a i r s program f o r each s u b r e g i o n o f 5. Second, i n -
A
s t e a d of j u s t one a d d i t i o n a l p o i n t
5,
we c o u l d u s e a whole c o l l e c t i o n{ j l ,
. . . , jv}
t o b u i l d a program of t h e t y p e (3.29).
A l l o f t h i s i s d e s c r i b e d i n d e t a i l i n [ l o ] f o r t h e c a s e when o n l y h i s s t o c h a s t i c b u t can e a s i l y be gener- a l i z e d t o t h e c a s e h and T s t o c h a s t i c .When o n l y q i s s t o c h a s t i c , we c o n s i d e r a d u a l v e r s i o n o f ( 3 . 2 9 ) , v i z . f o r
A A
e v e r y q = < E E and q = < E co E , l e t
such t h a t crA +
GT
I cA
w i t h
6
r [ 0 , 1 ] . The same a r g u m e n t s a s above w i t h5
=<, b u t r e l y i n g t h i s t i m e on t h e d u a l [ l l ] o f p r o b l e m ( 1 . 6 ) , l e a d t od -
(3.32) E{p ( c , c ) } z c x O + ~ ( x O ) : = i n f ( c * + i l )
.
4. DISCRETIZATION OF T H E PROBABILITY MEASURE P THROUGH CONDITIONAL EXPECTATIONS
J e n s e n ' s i n e q u a l i t y f o r convex f u n c t i o n s i s t h e b a s i c t o o l t o o b t a i n lower bounds f o r E when f ( x , 0) i s convex o r u p p e r bounds when E i s concave. Here, i t
f f
l e a d s t o t h e u s e o f ( m o l e c u l a r ) p r o b a b i l i t y measures c o n c e n t r a t e d a t c o n d i t i o n a l e x p e c t a t i o n p o i n t s . I n t h e c o n t e x t o f s t o c h a s t i c programming t h i s was f i r s t done by Madansky [12] and f u r t h e r r e f i n e d by K a l l [13] and Huang, V e r t i n s k y and Ziemba [14]
.
4.1 PROPOSITION. Let
S = {S R,
R = l ,...,
v )be a p a r t i t i o n o f
Z,w i t h
<'
= ~ { < ( w )I
s')and
pR = P [<(w) t s'].
Suppose f i r s t t h a t
< w f ( x , < )i s convex. Then
I f <I+ f (x, <)
i s concave, then
PROOF.
Follows from t h e i t e r a t e d a p p l i c a t i o n o f J e n s e n ' s i n e q u a l i t y : f (x, E{<(w) )) 5 E{f ( x , < ( w ) ) ) when f (x, 0) i s convex; c o n s u l t [15]. 0
4.4 APPLICATION.
C o n s i d e r t h e s t o c h a s t i c program w i t h r e c o u r s e w i t h o n l y h and T s t o c h a s t i c . With S = {S R,
R = l ,...,
v ) a p a r t i t i o n o f Z and f o r R = l ,...,
V , l e tR R R
5'
= (h yT ) = E{(h(w),T(w))1
S )and p = P[<(w) E St]
.
A s f o l l o w s from (1.11) and ( 4 . 2 ) , we o b t a i n Rand t h u s i f
- -
V R
zV = i n f n l cx +
1
pRQ(x,< ) I A x = b , x t O x t RI
R = 1where
R R R
Q ( x , S ) = i n f y r R n 2 [ q y
1
Wy=h-
T x , y s ~ ],
we have t h a t
z V I z* = i n f [ c x + Q ( x )
I
A x = b , x s o ].
Each zV i s t h u s a lower bound f o r t h e o p t i m a l v a l u e o f t h e s t o c h a s t i c program.
(An a l t e r n a t i v e d e r i v a t i o n o f ( 4 . 5 ) r e l y i n g on t h e d u a l o f t h e r e c o u r s e problem t h a t d e f i n e s Q(x,c) a p p e a r s i n [16] .)
v R
4.6 CONVERGENCE. Suppose S = {S
,
R = l ,. . . ,
v ) f o r v = l, . . . ,
a r e p a r t i t i o n s o f E w i t hsV
csV+'
and c h o s e n s o t h a t t h e Pv, v = l ,. . .
converge i n d i s t r i b u t i o nt o P . The P a r e t h e ( m o l e c u l a r ) p r o b a b i l i t y d i s t r i b u t i o n s t h a t a s s i g n p r o b a b i l i t y v
R R
p, = P[c(w) E S ] t o t h e e v e n t [((w) =
5
] wherece
i s t h e c o n d i t i o n a l e x p e c t a t i o n ( w i t h r e s p e c t t o P) o f c ( * ) g i v e n t h a t ((w) E S R.
The e p i - c o n v e r g e n c e of t h e {Qv, v = l ,. . .
) t oQ,
w i t h t h e accompanying convergence o f t h e s o l u t i o n s , f o l l o w s from Theorem 2 . 8 , whereTo make u s e o f t h e s e r e s u l t s we need t o d e v e l o p a s e q u e n t i a l p a r t i t i o n i n g s c h e m e f o r , i . e . given a p a r t i t i o n S o f v E how s h o u l d i t be r e f i n e d s o a s t o i m - prove t h e a p p r o x i m a t i o n t o Q a s much a s p o s s i b l e . P . K a l l h a s a l s o worked o u t v a r i o u s r e f i n e m e n t schemes [17]
.
4.7 I M P L E M E N T A T I O N . S t o c h a s t i c programs w i t h simple recourse, w i t h h s t o c h a s t i c , q and T a r e f i x e d . R e c a l l t h a t f o r a s t o c h a s t i c program w i t h s i m p l e r e c o u r s e
Y
t a k e s on t h e form:
f
m 2where
Ei
= hi and, a s f o l l o w s from ( 3 . 2 1 ) ,i i ( x i Y E i ) = q i ( h i - x i ) + i f h . 2 ~ i i ' q Y ( x i - h i ) i f h . 5 ~
i i '
4.8. Figure: The f u n c t i o n
ii (xi
, a )Let [ n . , B i ] b e t h e s u p p o r t o f t h e r e a l i z a t i o n s o f h i ( - ) , p o s s i b l y an unbounded
1
i n t e r v a l . I f we a r e o n l y i n t e r e s t e d i n a lower bound f o r
Y
t h a t approximates i tA
a s c l o s e l y a s p o s s i b l e a t t h e p o i n t X , t h e n t h e o p t i m a l p a r t i t i o n i n g o f [ a . , P . ]
1 1
i s g i v e n by
1 A 2 - A
S . = [cxi,xi) and
1 Si
-
[xi,Bi1.
In t h i s way t h e a p p r o x i m a t i n g f u n c t i o n
Y
a t a k e s on t h e form:i
where f o r R=1,2,
R R
h i = E { h i ( w ) l S } and Pig = P[hi(w) E S R
1 ,
and
h.
= E{hi (w)1 .
Note t h a t1
J A
h . (w) 1
>xi
a A
Thus Y. I Y . w i t h e q u a l i t y h o l d i n g f o r X . 1 1 1 < a i ,
x i > Bi
and a tx i = x i .
4.9. F i g u r e : The f u n c t i o n ya
I f t h e i n t e r v a l [ a .
,
Bi] h a s a l r e a d y been p a r t i t i o n e d i n v i n t e r v a l s1
0 1 V - 1 V A R R + 1
{ [ a i = a i , a i ) ,
. . . ,
[ a i , a . 1 =Bi] } andxi
E [ a i , a i ).
Then a g a i n t h e o p t i m a lR A
s u b d i v i s i o n of t h e i n t e r v a l [a!,aP+') i n t o [ a ) and [ x . ,a!+1) y i e l d s an e x a c t
1 1 1
bound f o r Yi a t
;(i.
An a l t e r n a t i v e i s t o s p l i t t h e i n t e r v a l under c o n s i d e r a t i o n around such t h a t ;i t u r n s o u t t o be t h e c o n d i t i o n a l e x p e c t a t i o n o f t h e newi
r e g i o n . This would p r o v i d e a q u i t e good bound f o r Y . i n t h e neighborhood of 1
Xi
and t h i s would be v e r y u s e f u l i f t h e v a l u e o f
xi
i s n o t e x p e c t e d t o change much i n t h e n e x t i t e r a t i o n s .4.10 IMPLEMENTATION.
General recourse matrix
W , w i t h h s t o c h a s t i c ; q and T a r e f i x e d . The f u n c t i o ni s n o t s e p a r a b l e , i t i s convex and p o l y h e d r a l ( 1 . 1 1 ) . Note a l s o t h a t
h i+ ijl(x,h-x)
i s a s u b l i n e a r f u n c t i o n . Because o f t h i s we s h a l l s a y t h a t $ ( x , * ) i s sublinear w i t h r o o t a t X . We assume t h a t E c R~~ i s a r e c t a n g l e and t h a t we a r e g i v e n a p a r t i t i o n {S R
,
R = l ,...,
v ) i l l u s t r a t e d below.4.11. F i g u r e : P a r t i t i o n S =
{sl ,...,
S v1
o f EWe s h a l l t a k e i t f o r g r a n t e d t h a t t h e n e x t p a r t i t i o n of ! w i l l b e o b t a i n e d by s p l i t t i n g one o f t h e c e l l s S R
.
O t h e r p a r t i t i o n i n g s t r a t e g i e s may b e u s e d b u t t h i s s i n g l e c e l l a p p r o a c h h a s t h e a d v a n t a g e o f i n c r e a s i n g o n l y m a r g i n a l l y t h e l i n e a r program t h a t n e e d s t o b e s o l v e d i n o r d e r t o o b t a i n t h e l o w e r bound.( i ) Let u s f i r s t c o n s i d e r t h e c a s e when
x
r S R c!.
We p l a n t o s p l i t S' w i t h a h y p e r p l a n e c o n t a i n i n gx
and p a r a l l e l t o a f a c e o fs',
o r e q u i v a l e n t l y p a r a l l e l t o a h y p e r p l a n e bounding t h e o r t h a n t s . To do t h i s , we s t u d y t h e b e h a v i o r of h t + $ ( x , h ) on e a c h e d g e Ek o f t h e c e l l S R.
L e tT h i s i s a p i e c e w i s e l i n e a r convex f u n c t i o n . The p o s s i b l e s h a p e o f t h i s f u n c t i o n i s i l l u s t r a t e d i n F i g u r e 4.12 below; by
x
P we d e n o t e t h e o r t h o g o n a l p r o j e c t i o n o fx
on E k .4.12. Figure: The f u n c t i o n Bk on Ek
I f 0 i s l i n e a r on E k , i t means t h a t we cannot improve t h e approximation t o 0
k k
by s p l i t t i n g SR s o a s t o s u b d i v i d e E ~ . On t h e c o n t r a r y i f t h e s l o p e s o f O k a t t h e end p o i n t s a r e d i f f e r e n t , t h e n s p l i t t i n g SR s o a s t o s u b d i v i d e E would i m -
k
p r o v e t h e approximation t o
Y .
On t h e s u b d i v i d e d c e l l s , t h e r e s u l t i n g f u n c t i o n s Bk would be c l o s e t o , i f n o t a c t u a l l y , l i n e a r . Among a l l edges E k , we would t h e n choose t o p a r t i t i o n t h e c e l l SR s o a s t o s u b d i v i d e t h e edge E t h a t e x h i b i t s f o rk
8 t h e l a r g e s t d i f f e r e n c e o f s l o p e s a t t h e end p o i n t s . What we need t o know a r e k
t h e s u b g r a d i e n t s o f t h e f u n c t i o n
s R
a t each v e r t e x {h
,
s = l ,...,
r ) o f t h e c e l l S.
This i s o b t a i n e d by s o l v i n g t h e l i n e a r programs(4.13) f i n d .rr E R~~ such t h a t .rrW I q and w =.rr(h - x ) S i s maximized
S
f o r s = l ,
. .
. , r . The o p t i m a l .rr S i s a s u b g r a d i e n t o f $ ( x , * ) a t hS [ l , P r o p o s i t i o n 7 - 1 2 ] . From t h i s we o b t a i n t h e d i r e c t i o n a l s u b d e r i v a t i v e of $ ( x , - ) i n each c o o r d i n a t ed i r e c t i o n (which a r e t h e s l o p e s o f t h e f u n c t i o n s 8 ) ; t h e y a r e s i m p l y t h e compo- k
n e n t s o f t h e v e c t o r {.rri, S i = l ,
. . . ,
m21 .
We now c o n s t r u c t a s u b d i v i s i o n o f S' w i t h a h y p e r p l a n e p a s s i n g throughx
and o r t h o g o n a l t o t h e edge o f S' t h a t e x h i b i t s maximum s l o p e d i f f e r e n c e . I f t h e u n d e r l y i n g p r o b a b i l i t y s t r u c t u r e i s such t h a t t h e random v e c t o r h ( * ) i s t h e sum o f a few random v a r i a b l e s , such a s d e s c r i b e d by ( 1 . 1 9 ) , t h e c a l c u l a t i o n o f t h e d i r e c t i o n a l s u b d e r i v a t i v e s o f S ~ $ b ( x , < ) a g a i n b e g i n s w i t h t h e c a l c u l a t i o n o f t h e o p t i m a l s o l u t i o n o f t h e programs (4.13) each h being o b t a i n e d a s t h e map o f a v e r t e x o f S S' through t h e map (1.19).
To o b t a i n t h e s u b d e r i v a t i v e s , we a g a i n need t o u s e t h i s t r a n s f o r m a t i o n .( i i ) We now c o n s i d e r t h e c a s e when
~4 E.
T h i s t i m e we cannot always choose a h y p e r p l a n e p a s s i n g throughx
t h a t g e n e r a t e s a f u r t h e r s u b d i v i s i o n of some c e l l S R.
Even when t h i s i s p o s s i b l e , i t might n o t n e c e s s a r i l y improve t h e approximation, t h e f u n c t i o n < k $ ( x , < ) b e i n g l i n e a r on t h a t c e l l f o r example.I d e a l l y , one s h o u l d t h e n s e a r c h a l l c e l l s S R and each edge i n any g i v e n c e l l t o f i n d where t h e maximum g a i n c o u l d be r e a l i z e d . G e n e r a l l y , t h i s i s i m p r a c t i c a l . What a p p e a r s r e a s o n a b l e i s t o s p l i t t h e c e l l w i t h maximum p r o b a b i l i t y , on which
$ ( x , * ) i s n o t l i n e a r .
Concerning t h e implementation o f t h i s p a r t i t i o n i n g t e c h n i q u e , we a r e s e e k i n g t h e approximation t o
Y
which i s a s good a s p o s s i b l e i n t h e neighborhood o f X . We a r e t h u s working w i t h t h e i m p l i c i t assumption t h a t we a r e i n a neighborhood o f t h e o p t i m a l s o l u t i o n and t h a tx
w i l l n o t change s i g n i f i c a n t l y from one i n t e r - a c t i o n t o t h e n e x t . I f t h i s i s t h e c a s e , and t h e problem i s w e l l - p o s e d , t h e n wes h o u l d n ' t r e a l l y h a v e t o d e a l w i t h c a s e ( i i ) , s i n c e i t would mean t h a t t h e o p t i - ,ma1 t e n d e r
x
0 would b e s u c h t h a t we would c o n s i s t e n t l y u n d e r e s t i m a t e o r o v e r e s t i -mate t h e demand!
4.14 APPLICATION. C o n s i d e r ' t h e s t o c h a s t i c program w i t h r e c o u r s e w i t h o n l y q s t o c h a s t i c . With S = {S R
,
R = l ,...,
v ) a p a r t i t i o n o f Z , and f o r R = l ,...,
v , l e tand p R = P[<(w) E S R
1 .
AS f o l l o w s from ( 1 . 1 2 ) and ( 4 . 3 ) we h a v eThus, w i t h
v v R
Z = i n f x C R n l c x
[
+ R = l1
pRQ(x.S )I
Ax = b y x 2 whereR R
Q(x.
S 1
= i n f Y E R n 2 [ q YI
WY = h-Tx, Y 2 01,
we have t h a t
z v 2 z* = i n f [ c x + Q ( x )
I
Ax = b y x 2 01.
Each z v i s t h u s a n u p p e r bound f o r t h e o p t i m a l v a l u e o f t h e s t o c h a s t i c program.
4.15 :IMPLEMENTATION. The f u n c t i o n
i s p o l y h e d r a l and s u p l i n e a r . What changes from one
x
t o t h e n e x t a r e t h e s l o p e s o f t h i s f u n c t i o n , s o we c a n n o t u s e t h e p r e s e n tx
a s a g u i d e f o r t h e d e s i g n o f t h e a p p r o x i m a t i o n . One p o s s i b i l i t y i n t h i s c a s e i s t o s i m p l y s u b d i v i d e a c e l l o f t h e p a r t i t i o n w i t h maximum p r o b a b i l i t y .5. DISCRETE P R O B A B I L I T Y MEASURES W I T H EXTREMAL SUPPORT
The maximum o f a convex f u n c t i o n on a compact convex s e t i s a t t a i n e d a t an extreme p o i n t ; moreover, t h e f u n c t i o n v a l u e a t any p o i n t (of i t s domain) o b t a i n e d a s a convex computation o f extreme p o i n t s i s dominated by t h e same convex combin- a t i o n of t h e f u n c t i o n v a l u e s a t t h o s e extreme p o i n t s . These e l e m e n t a r y f a c t s a r e used i n t h e c o n s t r u c t i o n o f measures t h a t y i e l d upper bounds f o r t h e e x p e c t a t i o n f u n c t i o n a l E f .
5 . 1 PROPOSITION. Suppose c c t f ( x , c ) i s convex,
E
t h e support o f t h e random v a r i - able 5 ( * ) i s compact, and l e t e x t E denote t h e extreme points o f coE,
t h e convex h u l l o f.
Suppose moreover t h a t f o r a l l5,
v ( c , * ) i s a p r o b a b i l i t y measure de- fined on ( e x t E, E) w i t h E t h e Borel f i e l d , such t h a tj e v(5,de) =
5
¶e x t ::
and
w
"
v(5(w) ¶A)i s measurable for a l l A E
E .
Then ( 5 . 2 )e x t =.
where A i s t h e p r o b a b i l i t y measure on
E
defined byPROOF. The c o n v e x i t y o f f ( x , - ) i m p l i e s t h a t f o r t h e measure v a s d e f i n e d above f ( x , ~ ) 5
J'
f ( x , e ) v ( ~ , d e ).
e x t
-
zS u b s t i t u t i n g E ( * ) f o r